Task 003: Neuro-inspired Algorithms and Theory

Abstract:

Edge-computing requires the development of simple low-power systems capable of performing complex computation. Hyperdimensional (HD) computing is a computational framework inspired by the observation that the brain represents mental events by the simultaneous firing of a pattern of dispersed neurons. HD has been successfully applied to a range of machine learning problems.  Classification tasks are performed using large vectors, containing thousands of dimensions, to store a representation of the data as a holographic distributed pattern. Simple computational operations exploit the statistical properties of high dimensional vector spaces to perform energy-efficient and highly parallelalizable classification that is robust to noise or error. Training data is encoded into hypervectors which can be bundled to create class hypervectors, naturally allowing for incremental one-shot learning. One method of encoding the hypervectors uses the non-linear dynamical behavior of Cellular Automaton (CA) and the principles of Reservoir Computing (RC). Input is mapped onto the initial CA state and then projected to high-dimensional space by repeated application of the rule. The space-time volume of the automatons form reservoirs which can be combined using HD operations. Different encodings are best able to capture particular features of the input data. We therefore propose to increase the accuracy by using a set of two or more CA rules to build a classifier with multiple encodings. After training, the model can be binarized for efficient low-cost classification using hamming-distance similarity. By combining complementary CA rules a simple power-efficient architecture can be developed that improves on the classification accuracy of current state-of-the-art HD computing techniques.

 

Bio:

Neil Babson is a PhD student in Computer Science. He previously received an M.S. in Computer Science from Portland State as well as B.S. degrees in Mathematics and Physics. Neil’s interests include crochet and piano.